🧩️ The Missing Piece: Why AI Might Need Its Own Operating System
/dev/agents and Gemini 2.0 point to the future of work
Hiya,
"Before Android existed, we could all see the promise of mobile, but as a developer, it was just too hard to build anything on mobile," says David Singleton.
He should know - as one of the key architects of Android, he helped solve that problem. Now, as CEO of /dev/agents, he's tackling a similar challenge with AI.
Two weeks ago, his startup raised $56M at a $500M valuation to build what they're calling an "operating system for AI agents." Yesterday, Google announced Gemini 2.0, heralding what they call "the agentic era" of AI.
Wait, what’s an agentic era again?
The "agentic era" of AI refers to a phase in artificial intelligence where autonomous AI systems, or agents, are no longer merely tools responding to direct human commands but become proactive collaborators that make decisions, plan actions, and execute tasks with minimal human intervention.
In theory.
This next evolutionary step won’t happen without some human intervention.
The /dev/agents team, which includes other Android veterans and former leaders from Meta and Stripe, sees the exact same pattern emerging with AI that they saw with mobile fifteen years ago. The fact that Google, which just announced arguably the most advanced AI model in the world, had already bet on this infrastructure play through their investment arm CapitalG tells us something important about where AI is heading.
🔍 CLEAR VIEW: History Doesn't Repeat, But It Rhymes
Think back to 2008. We had powerful mobile hardware. We had semi-exciting apps. What we didn't have was a platform that made it all work together reliably. Android and iOS changed that.
Now look at 2024. We have AI that can:
Write and optimise marketing campaigns
Analyse customer behavior in real-time
Generate and test ad creative
Manage bidding across platforms
Handle customer service inquiries
But try to get these AI tools to work together, and you hit a wall.
Want your analytics AI to automatically brief your content AI when it spots a trend? Want your bidding AI to coordinate with your creative AI when certain ads perform better? Want all of these systems to share context about what's working and what isn't?
Go swivel.
Right now, that's nearly impossible without custom development.
Let's look at what this means in practice, using the example of a digital marketer. They might start the day like this:
This looks familiar to me. That’s why I made it that way, I suppose.
A lot of time is spent switching between platforms and programs, pulling data and analysing it, then discussing what to do next. We do not work at the speed of data.
Now, let’s see how that flow might look in the agentic era:
Looks pretty sweet, huh?
My challenge here is that we’ve seen this all before. We’ve all heard the breezy promise that “technology will handle the drudgery”, which will somehow “liberate us to be strategic thinkers”.
Does that sound like the world we actually live in to you?
If we weren’t strategic thinkers already, some automation won’t change that overnight.
Moreover, we retain working structures built in the industrial revolution to maximise the units of production from each cog in the machine. It’s not particularly effective, but it’s the system we have.
Bosses will not suddenly let everyone mooch around like Beat poets waiting for divine inspiration to strike, while the AI does what we recognise today as “the real work”. One of those groups will be made redundant and I’ll eat my beret if it’s the robots.
And yet, I’m somewhat sanguine about what’s possible here. If we try to graft AI onto what we’re already doing, we won’t get full value from it. It requires structural changes and it’s plausible that an entirely new operating system is required.
🔮 THROUGH THE GLASS: What Makes An OS Special
The /dev/agents team, backed by Google's CapitalG and other major investors, isn't just building better AI tools. They're building the infrastructure that makes AI tools work together. Think about what Android provides:
Resource management
Security frameworks
Inter-app communication
Consistent user interfaces
Standard development tools
Now imagine the equivalent for AI agents:
Shared Memory When your marketing agent notices that CTR drop at 3 AM, it's not just seeing numbers - it has access to a shared context layer in the OS that knows:
Historical performance patterns
Current campaign objectives
Budget constraints
Brand guidelines
Previous optimisation attempts
It's like giving all your AI tools access to the same institutional knowledge.
Resource Management The OS doesn't just let agents run wild. When the marketing agent wants to run those small tests at 5 AM, the OS:
Checks permissions and budgets
Ensures tests won't interfere with other campaigns
Manages computational resources
Tracks and logs all actions
You get the safety of oversight with the speed of automation.
Does this alter my assessment about whether people will be somehow more vital in this new working relationship?
It can, if people adapt to the new way of working. These AI agents are not magical. They are not blessed with an innate knowledge of what makes a brand tick. They’re closer to a toddler with superpowers; capable of the extraordinary, but still in need of serious boundaries.
Marc Benioff, CEO of Salesforce, gave an example in an interview this week. He said that he recently suffered a leg injury. He had surgery and they sent him off with some painkillers and an instruction to come back and see the doctor in a month or two. One has to imagine that Mr Benioff has reasonable healthcare coverage, too.
As he noted, this isn’t perfect service. However, the doctors are making the best possible use of the labour resources at their disposal. The patient may have questions or minor complications, but they will only contact the doctor in more extreme circumstances. AI agents could help deliver an ongoing service to the patient, stepping in to help with those additional queries along the way.
The idea is not to do the doctor’s job for them, but rather to increase their productive capacity and improve the service they offer. It’s not the best example in the world, to be honest, and we must remember that Mr Benioff has a new service (Agentforce) to sell.
Nonetheless, we can apply the same thinking to our marketing example.
The AI agents are able to analyse campaign CTR during the night, test new creative assets, and send a report to the human marketer. Lovely stuff, but it’s not the end of the marketer’s task list. They still have 1000s of other things they could be doing to improve the company’s marketing effectiveness. If they know how to use AI agents, and they have the structures in place to let them collaborate with these agents effectively, the marketer can finally start making a dent in that to-do list.
In summary
The agentic era isn’t just about automating repetitive tasks. It’s about changing the way we think about work.
Google announcing Gemini 2.0's agentic capabilities so soon after we learned about their investment in /dev/agents tells us something important: Even as they push the boundaries of what AI can do, they recognise that capability without coordination isn't enough.
It's not just about making AI more powerful - it's about making it more usable. Just as the smartphone age needed Android and iOS, Google is betting that AI needs something similar. Expect announcements from Google’s rivals in the coming months, too.